MSU TECHNICAL REPORT, MSU-CSE-10-14, AUGUST 2010 (SUBMITTED TO IEEE TIFS) 1 Latent Fingerprint Matching: Utility of Level 3 Features Qijun Zhao, Jianjiang Feng, and Anil K. Jain, Fellow, IEEE Abstract Automatic fingerprint identification systems (AFIS) have for a long time used only minutiae for fingerprint matching. But minutiae are only a small subset of fingerprint detail routinely used by latent examiners for fingerprint matching. This has generated a lot of interest in extended feature set (EFS) with the aim of narrowing down the gap between the performance of AFIS and latent examiners. Level 3 features constitute the most significant subset of extended features. Studies on level 3 features have reported significant improvement in the fingerprint recognition accuracy. However, these studies were based either on live-scan fingerprints or full (rolled or slap) fingerprints. As a result, the conclusions of these studies cannot be extended to latent fingerprints, which are characterized by small size, poor image quality, and severe distortion compared to full fingerprints. In this paper, we study the utility of level 3 features, including pores, dots, incipient ridges, and ridge edge protrusions, for latent matching. Automatic algorithms for extracting and matching these features are proposed. While most existing level 3 feature matching algorithms only consider the locations of features, the proposed method utilizes the topological relationship between level 3 and level 2 features, and is thus robust to nonlinear distortion and has high discriminative capability. Given the proposed algorithms and operational latent fingerprint databases, we identify the challenges in using level 3 features, and show the potential of level 3 features in improving latent matching accuracy. Further, by using simulated partial fingerprints, we highlight that level 3 features can indeed improve latent matching accuracy when i) level 3 features can be reliably extracted in both latent and full fingerprints and ii) latent fingerprints have only a small number of minutiae or the minutiae match scores are low. With the increasing adoption of 1000ppi fingerprint scanners in law enforcement agencies, it is becoming feasible and desirable to incorporate level 3 features into AFIS. We believe that the proposed algorithms and analysis will be useful in the design and development of next generation AFIS. Index Terms Latent fingerprint matching, extended features, level 3 features, forensics ✦
30
Embed
MSU TECHNICAL REPORT, MSU-CSE-10-14, AUGUST 2010 ...biometrics.cse.msu.edu/Publications/Fingerprint/ZhaoFengJain_Late… · MSU TECHNICAL REPORT, MSU-CSE-10-14, AUGUST 2010 (SUBMITTED
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
MSU TECHNICAL REPORT, MSU-CSE-10-14, AUGUST 2010 (SUBMITTED TO IEEE TIFS) 1
Latent Fingerprint Matching:
Utility of Level 3 FeaturesQijun Zhao, Jianjiang Feng, and Anil K. Jain, Fellow, IEEE
Abstract
Automatic fingerprint identification systems (AFIS) have for a long time used only minutiae
for fingerprint matching. But minutiae are only a small subset of fingerprint detail routinely used
by latent examiners for fingerprint matching. This has generated a lot of interest in extended
feature set (EFS) with the aim of narrowing down the gap between the performance of AFIS
and latent examiners. Level 3 features constitute the most significant subset of extended features.
Studies on level 3 features have reported significant improvement in the fingerprint recognition
accuracy. However, these studies were based either on live-scan fingerprints or full (rolled or
slap) fingerprints. As a result, the conclusions of these studies cannot be extended to latent
fingerprints, which are characterized by small size, poor image quality, and severe distortion
compared to full fingerprints. In this paper, we study the utility of level 3 features, including pores,
dots, incipient ridges, and ridge edge protrusions, for latent matching. Automatic algorithms for
extracting and matching these features are proposed. While most existing level 3 feature matching
algorithms only consider the locations of features, the proposed method utilizes the topological
relationship between level 3 and level 2 features, and is thus robust to nonlinear distortion and has
high discriminative capability. Given the proposed algorithms and operational latent fingerprint
databases, we identify the challenges in using level 3 features, and show the potential of level 3
features in improving latent matching accuracy. Further, by using simulated partial fingerprints, we
highlight that level 3 features can indeed improve latent matching accuracy when i) level 3 features
can be reliably extracted in both latent and full fingerprints and ii) latent fingerprints have only a
small number of minutiae or the minutiae match scores are low. With the increasing adoption of
1000ppi fingerprint scanners in law enforcement agencies, it is becoming feasible and desirable to
incorporate level 3 features into AFIS. We believe that the proposed algorithms and analysis will
be useful in the design and development of next generation AFIS.
Algorithm 1 describes the ridge correspondence establishment. Suppose a set of mated minutiae are
found betweenFq andFt by the minutiae matcher. Fig. 9 shows an example latent and its mated rolled
fingerprint in ELFT-EFS-PC. There are three pairs of mated minutiae in them. From each pair of mated
minutiae, several pairs of candidate aligned ridges can be obtained from the ridges associated with the
two minutiae in the pair. A candidate aligned ridge pair is defined asCR = {RSP1, PR1;RSP2, PR2},whereRSP1 andRSP2 are the two candidate ridges (or ridge segments) represented by their sampling
points and the first sampling points on them are assumed to be matched (here, the sampling points
corresponding to the mated minutiae) andPR1 andPR2 are the parent ridges from which this candidate
aligned ridge pair is generated. For the candidate aligned ridge pairs generated from mated minutiae,
the parent ridges are set to 0, which means they have no parentridges. The parent ridges will be used
in Inter-Ridge propagation to ensure that only the samplingpoints neighboring to the parent ridges are
matched during the propagation. For example, for the mated minutiae pair{Fq.M1, Ft.M1} in Fig. 9(a),
each of the three ridges associated withFq.M1 is paired with the corresponding ridge associated with
MSU TECHNICAL REPORT, MSU-CSE-10-14, AUGUST 2010 (SUBMITTED TO IEEE TIFS) 14
Ft.M1, resulting in three pairs of candidate aligned ridges. Notethat if a ridge ending is mated with a
ridge bifurcation, we will have two pairs of aligned ridges,while if two ridge endings are mated, we
will get only one pair of aligned ridges.
From each of these aligned ridge pairs, the two ridges in the pair are first compared by the Intra-
Ridge matching procedure. If the two ridges can be matched (i.e. more than four sampling points are
matched between them), the Inter-Ridge propagation procedure is invoked to match the remaining ridges
in the two fingerprints based on the mated sampling points on the two ridges. After all the aligned ridge
pairs have been considered, the ridge correspondences obtained from the one which gives the highest
similarity between the ridges in the two fingerprints are taken as the final result. Next, we introduce
the two main procedures, Intra-Ridge matching and Inter-Ridge propagation, which are involved in ridge
correspondence establishment.
3.2.1 Intra-Ridge Matching
Given a candidate aligned ridge pairCR = {RSP1, PR1, RSP2, PR2}, Intra-Ridge matching is used to
find the corresponding sampling points on the two aligned ridges (or ridge segments). This is essentially a
string matching problem given that the first sampling pointsin RSP1 andRSP2 are matched. We employ
the dynamic programming technique [21] to find the longest sequence of mated sampling points on the
two ridges,MRSP = {RSPm1 , RSPm
2 }, such that (i) the indices of mated sampling points monotonously
increase in bothRSPm1 andRSPm
2 , (ii) changes between indices of adjacent mated sampling points are
less than 3 (i.e. no more than 3 sampling points can be skippedduring matching), and (iii) ifPRi 6= 0,
all the mated sampling points inRSPmi should havePRi as their neighboring ridges (i = 1, 2). In our
implementation, two mated sampling points should satisfy i) the absolute difference between the distances
from them to the first sampling points is below a given threshold (i.e. 10 pixels), and ii) the absolute
difference between the ridge curvatures at them is also below a given threshold (i.e. 15 degrees). We
measure the distance between two sampling points on a ridge by using the absolute difference between
their indices, which is similar to geodesic distance. The ridge curvature at a sampling point is measured
by the change in local ridge orientation at the point with respect to the ridge orientation at the first
sampling point.
Given the mated points between two ridges, the similarity between the ridges is computed as follows.
Because short ridges are mostly unreliable, if there are fewer than 4 mated points between two ridges,
we discard them. Otherwise, we further examine the neighboring ridge structures of the mated sampling
points. Letnmsp = |MRSP | be the number of mated sampling points found onRSP1 and RSP2.
MSU TECHNICAL REPORT, MSU-CSE-10-14, AUGUST 2010 (SUBMITTED TO IEEE TIFS) 15
(a) (b)
(c) (d)
Fig. 9. Ridge correspondence establishment results for the latent L177B and its mated exemplar
in ELFT-EFS-PC. (a) Three pairs of mated minutiae are found between the latent and its exemplar.
(b) The ridges Fq.R1 and Ft.R1 associated with the mated minutiae pair {Fq.M1, Ft.M1} are
matched by Intra-Ridge matching. (c) The mated ridges found at an intermediate step as the
procedure Inter-Ridge propagation proceeds from the mated ridges in (b). (d) The final mated
ridges between the two fingerprints obtained by the proposed method. Corresponding ridges are
marked by the same color.
For all the mated sampling points inRSPmi (i = 1, 2), we examine on left-hand and right-hand sides,
repsectively, if the neighboring ridges of each two adjacent sampling points are different or not, resulting
in two feature vectors,NRli ∈ {0, 1}nmsp−1 and NRr
i ∈ {0, 1}nmsp−1, in which ‘0’ means same ridge
and ‘1’ means different ridges.NRl1 andNRr
1 are then compared withNRl2 andNRr
2, respectively, and
the number of the same entries between them is counted, denoted asnlNR and nr
NR for the left-hand
and right-hand sides, respectively. The similarity between the neighboring ridge structures of the mated
sampling points on the two ridges is then calculated by
sN = 0.5 × nlNR
nmsp − 1+ 0.5 × nr
NR
nmsp − 1. (3)
MSU TECHNICAL REPORT, MSU-CSE-10-14, AUGUST 2010 (SUBMITTED TO IEEE TIFS) 16
If the mated sampling points on the two ridges have very low similarity between their neighboring ridge
structures, they are also discarded. Fig. 9(b) shows the Intra-Ridge matching results for the ridgesFq.R1
andFt.R1 associated with the mated minutiaeFq.M1 andFt.M1 in Fig. 9(a).
3.2.2 Inter-Ridge Propagation
Given a set of mated sampling points on the two ridges found bythe Intra-Ridge matching procedure,
the Inter-Ridge propagation procedure, as sketched in Algorithm 2, matches all the remaining ridges. A
queue (denoted asQ) is constructed to store the candidate aligned ridge pairs.The queue is initialized by
generating candidate aligned ridge pairs from each pair of mated sampling points. The candidate aligned
ridge pairs are the neighboring ridges on the correspondingsides of the mated sampling points.
Algorithm 2 Inter-Ridge PropagationInput: MRSP : Mated sampling points on two ridges inFq andFt; Rq, Rt: Ridges inFq andFt
Output: sr: Similarity between ridges inFq andFt; MR: Mated ridge pairs betweenFq andFt, and
corresponding sampling points on them
1: MR←MRSP
2: Initialize the queue of candidate aligned ridge pairs,Q, based on MRSP
3: while Q is not emptydo
4: Retrieve the first candidate aligned ridge pair inQ: CR
5: mrsp← IntraRidgeMatch(CR)
6: if |mrsp| > 4 then
7: Appendmrsp to MR
8: Generate new candidate aligned ridge pairs based onmrsp
9: Push the new candidate aligned ridge pairs intoQ
10: end if
11: end while
12: Calculate the similarity between the ridges inFq andFt: sr
After the initialization ofQ, we start the main loop of the Inter-Ridge propagation procedure to compare
the ridges in each of the candidate aligned ridge pairs inQ until Q is empty. The first candidate inQ
is popped out and matched by the Intra-Ridge matching procedure. If more than four mated sampling
points are established, new candidate aligned ridge pairs are generated and pushed intoQ. WhenQ is
empty, the Inter-Ridge propagation procedure terminates with a set of mated ridge pairs as well as the
MSU TECHNICAL REPORT, MSU-CSE-10-14, AUGUST 2010 (SUBMITTED TO IEEE TIFS) 17
corresponding mated sampling points. Figs. 9(c) and 9(d) shows the mated ridge pairs found between the
two example fingerprints as the procedure Inter-Ridge propagation proceeds from the mated ridge pairs
shown in Fig. 9(b).
Let s̄N , d̄loc, and d̄ori be the average similarity between neighboring ridge structures of all the mated
ridge pairs, the average location displacement and the average orientation difference between all the
mated sampling points, respectively. LetnMRSP andnRSP denote the total number of mated sampling
points on the ridges in the two fingerprints and the total number of sampling points on ridges in the
query fingerprint, respectively. Then the similarity between the ridges in the two fingerprints is defined
as
sr = 0.3 × s̄N + 0.2× 10− d̄loc
10+ 0.2× 15− d̄ori
15+ 0.3× nMRSP
nRSP. (4)
3.3 Pore Matching
Once the correspondences between ridges are obtained, the level 3 features can be matched along the
mated ridges. To implement this, we need to first associate the level 3 features with the ridges. In this
section, we discuss the matching of pores; the matching of DIP features will be discussed in the next
section. Recall that pores in a fingerprint are all located onridges. Hence, for each pore, we find the
closest ridge to it and its projection point on this ridge. The pores on the same ridge are then grouped
together and ordered along the ridge tracing direction.
Given a pair of mated ridges, the correspondences between pores on these two ridges are found using
the following method. For each pore on a ridge,POR1, we first find its closest sampling pointSP1 on the
ridge (denote the ridge asR1). If SP1 does not have a mated sampling point, thenPOR1 does not have
any mated pores; otherwise, we find the nearest pore,POR2, to the mated sampling point,SP2, of SP1.
The location displacementdi betweenPORi andSPi (i = 1, 2) is calculated as the difference between
the sampling indices of the projection point ofPORi andSPi. The location displacement between the
two pores,POR1 andPOR2, is then defined asdloc = |d1−d2|. If dloc is smaller than a given threshold
(i.e. 10 pixels for 1000ppi fingerprint images),POR1 is mated withPOR2. After all the pores in the
latent are examined, we get the mated pores between the two fingerprints. Fig. 10(a) shows the mated
pores obtained between a latent and its mated exemplar.
To calculate the pore match score, we compare the neighboring ridge structures and pore distribution
of the mated pores on each pair of mated ridges (recall that the pores on a ridge are ordered, so are
the mated pores). A comparison of neighboring ridge structures is the same as being described for
mated sampling points on ridges (see Section 3.2.1). As for the neighboring pore distribution, if two
MSU TECHNICAL REPORT, MSU-CSE-10-14, AUGUST 2010 (SUBMITTED TO IEEE TIFS) 18
mated pores both have a neighboring pore on its left-hand side or right-hand side neighboring ridge, the
location displacement between the neighboring pores is calculated. Letdl,iNP be the location displacement
between the neighboring pores on the left-hand side of theith mated pores on the mated ridges, anddr,iNP
the location displacement between the neighboring pores onthe right-hand side of them. The similarity
between the neighboring ridge structures and pore distribution of the mated pores in the two ridges can
be then calculated as
sN = 0.4× nlNR
nmp − 1+ 0.4× nr
NR
nmp − 1+ 0.1× Σ
nlNP
i=1 (10− dl,iNP )
10× nlNP
+ 0.1× Σnr
NP
i=1 (10 − dr,iNP )
10 × nrNP
, (5)
where nmp is the number of mated pores on the two ridges andnlNP and nr
NP are the number of
cooccurring neighboring pores on the left-hand and right-hand sides, respectively. Finally, the pore match
score between the two fingerprints is defined as
sPOR = 0.8× sr + 0.2× (0.3 × s̄N + 0.3× 10− d̄loc
10+ 0.4× nMP
nP), (6)
where s̄N is the average similarity between neighboring ridge structures and pore distribution of all
the mated pores on the mated ridge pairs,d̄loc is the average location displacement between all the
mated pores, andnMP andnP denote the total number of mated pores and the number of poresin the
query latent fingerprint, respectively. It is worth mentioning that the above match score measures the
similarity between fingerprints by considering not only thelocation displacement between mated pores
and the number of mated pores, but also the consistency of theridge structures and feature distribution
surrounding the mated pores, whereas existing methods [10][12] consider only the location displacement
or the number of mated pores.
3.4 DIP Matching
The matching of DIP features is also constrained along matedridges. Unlike pores, DIP features reside
on valleys rather than ridges. Therefore, we associate eachDIP feature with two ridges that are on the
left-hand and right-hand sides of the valley on which it resides. Given a DIP featureDIP1 which is
associated with two ridgesR11 andR12, the nearest sampling point to its projection on the ridgeR11
is first found, denoted asSP1. If SP1 does not have mated sampling points, thenDIP1 does not have
mated DIP features; otherwise, the nearest DIP feature to the mated sampling pointSP2 on the mated
ridge R21 of R11 is found, denoted asDIP2. Let di be the location displacement betweenDIPi and
SPi (i = 1, 2), then the location displacement betweenDIP1 andDIP2 is dloc = |d1 − d2|. Let R22 be
the other ridge associated withDIP2. DIP1 andDIP2 are mated DIP features only if i)dloc ≤ 10 and
MSU TECHNICAL REPORT, MSU-CSE-10-14, AUGUST 2010 (SUBMITTED TO IEEE TIFS) 19
(a) (b)
Fig. 10. Example level 3 feature matching results. (a) Mated pores in the latent shown in Fig.
3(b) and its mated exemplar. Corresponding pores are marked by the same color. (b) Mated DIP
features in the latent shown in Fig. 9(a) and its mated exemplar.
ii) R12 andR22 are mated ridges. After enumerating all the DIP features inFq, we obatin the mated DIP
features. Fig. 10(b) shows the obtained mated DIP features in a latent and its mated exemplar.
Let d̄loc be the average location displacement between all the mated DIP features andnMDIP and
nDIP be the total number of mated DIP features and the number of DIPfeatures in the query latent
fingerprint, respectively. The DIP match score between the two fingerprints is then defined as
sDIP = 0.8 × sr + 0.2 × (0.3× 10− d̄loc
10+ 0.7 × nMDIP
nDIP). (7)
4 EXPERIMENTS
4.1 Databases
Two latent databases were used in this study. One is the ELFT-EFS-PC database [16], which has 242
1000 ppi latent fingerprints (most of them are from the same source as the latents in NIST SD27) with
1000 ppi mated full prints. The level 3 features in these latents have already been manually marked by
latent examiners. The background database consists of 4,180 1000 ppi fingerprint images, which were
collected from the mated fingerprints of the latents and the “B” session fingerprint images in the NIST
SD29 and the NIST SD30 datasets. The second latent database was collected at West Virginia University
(WVU). It has 127 latents in which the level 3 features have been manually marked. While these latents
are at 1000 ppi, the full fingerprints in the background database are only at 500 ppi. In an earlier study
[18], we investigated the utility of pores in the context of varying fingerprint image quality and resolution
by using the rolled ink fingerprint images in NIST SD30 and a commercial minutiae matcher (VeriFinger
MSU TECHNICAL REPORT, MSU-CSE-10-14, AUGUST 2010 (SUBMITTED TO IEEE TIFS) 20
0 10 20 30 40 50 60 70 80 900
100
200
300
400
500
600
700
800
900
No. of Minutiae
No.
of P
ores
ELFT−EFS−PCWVU
(a)
0 10 20 30 40 50 60 70 80 900
5
10
15
20
25
30
35
40
45
50
No. of Minutiae
No.
of D
IP
ELFT−EFS−PCWVU
(b)
Fig. 11. The number of features in the latents in the ELFT-EFS-PC and WVU databases. (a)
Number of pores vs. Number of minutiae. (b) Number of DIP features vs. Number of minutiae.
[19]). It was reported that automatic pore extraction and the resulting matching accuracy are significantly
affected by fingerprint image quality. Further, it is only athigh resolution (1000 ppi) and for good quality
fingerprint images that the pores can improve the fingerprintverification accuracy, and even then only
marginally. Hence, the WVU database is not suitable for studying the utility of level 3 features in latent
matching (although many latents in it have large number of pores, the mated rolled images at 500 ppi
do not have a sufficient number of pores), and we simply used itfor reporting the statistics of level 3
features in latents.
4.2 Statistics of Level 3 Features in Latents
The statistics of level 3 features in the latents in the ELFT-EFS-PC and WVU databases have been
collected based on the manual markup data. Fig. 11 shows the number of level 3 features (i.e. pores and
DIP features) with respect to the number of minutiae. It can be seen that there is a large variance in the
number of level 3 features across different latent fingerprints. In ELFT-EFS-PC, very few of the latents
(only 6 of 242 latents) have any pores. Fig. 12 shows three example latents in ELFT-EFS-PC, which
have 91, 17, and 0 pores marked by the latent exminers, respectively. Given such a small number of
latents in ELFT-EFS-PC which have pores, pore matching is not likely to improve the latent matching
accuracy on this database. Since no public domain latent database is suitable for studying the utility of
pores, we constructed a simulated partial fingerprint database and investigated the effectiveness of pores
on that dataset (see Section 4.5). On the other hand, many (79out of 242) latents in ELFT-EFS-PC do
MSU TECHNICAL REPORT, MSU-CSE-10-14, AUGUST 2010 (SUBMITTED TO IEEE TIFS) 21
Fig. 12. Example latents in ELFT-EFS-PC with markup pores. (a) 91 pores, (b) 17 pores, and (c)
0 pores.
have DIP features. Therefore, we will use the ELFT-EFS-PC database for studying the effectiveness of
DIP features.
4.3 Feature Detection Accuracy
Sixty partial fingerprints (320×240 pixels) were cropped from the 1000 ppi rolled ink fingerprint images
(∼1500×1500 pixels) in NIST SD30. The pores, dots, incipient ridges, and ridge edge protrusions in
these sixty partial fingerprint images were manually markedfor the purpose of evaluating their automatic
detection accuracy. In order to study the impact of image quality on the automatic detection accuracy,
the 60 partial fingerprint images were divided into two quality groups (good and bad) according to their
image quality evaluated by the method in [22]. Two pore detection methods were considered, i.e. the
proposed method and the method in [10]. These two methods differ in that the proposed method conducts
filtering in the spatial domain using an anisotropic model, whereas the method in [10] applies filtering
in the frequency domain with an isotropic model. Table 1 gives the average pore detection accuracy
along with the standard deviation of the two methods on the ground truth dataset;Rt, the true detection
MSU TECHNICAL REPORT, MSU-CSE-10-14, AUGUST 2010 (SUBMITTED TO IEEE TIFS) 22
TABLE 1
Average pore detection accuracy and standard deviation
Method Proposed Method in [10]
Quality Good Bad Good Bad
Rt(%) 73 ± 9.5 67 ± 14.6 69 ± 16.0 63 ± 12.5
Rf (%) 20 ± 12.2 30 ± 14.4 27 ± 11.7 40 ± 18.1
TABLE 2
Average DIP detection accuracy and standard deviation
Quality Good Bad
Nm 6 ± 5.6 12 ± 8.5
Ns 4 ± 3.6 10 ± 7.0
rate, is defined as the ratio of the number of true detected pores to the total number of ground truth
pores, andRf , the false detection rate, is defined as the ratio of the number of falsely detected pores to
the total number of detected pores. These results show that the detection accuracy of both the methods
degrades as the fingerprint image quality goes down. As the fingerprint image quality changes from good
to bad, the true detection rate decreases by about 5%, and thefalse detection rate increases more by
about 10%. According to the standard deviation of the detection accuracy, in general, the automatic pore
detection methods are more robust on good quality fingerprint images than on bad quality fingerprint
images. Table 2 presents the detection accuracy of the proposed DIP detection method ([10] did not
present a DIP feature extractor), whereNm andNs denote the numbers of missing features and spurious
features, respectively. Note that the average numbers of observable DIP features on good and bad quality
fingerprint images in the ground truth dataset are 12 and 27, respectively. Again, poor quality fingerprint
images cause more missing features as well as more spurious features. These results show that improving
the quality of full fingerprint images in the background database is very important for the effectiveness
of level 3 features in latent fingerprint matching. We will further demonstrate this in the next section.
4.4 Latent Fingerprint Matching
We have evaluated the latent matching performance with pores and DIP features on the 242 latents in
ELFT-EFS-PC. Features (including minutiae, ridge skeletons, pores, and DIP) were manually marked in
MSU TECHNICAL REPORT, MSU-CSE-10-14, AUGUST 2010 (SUBMITTED TO IEEE TIFS) 23
(a) (b)
Fig. 13. Poor quality of exemplar fingerprints significantly degrades the utility of level 3 features.
(a) Latent L030G in ELFT-EFS-PC, and (b) its mated full fingerprint image. Nine mated minutiae
are found between them. But none of the 91 pores marked in the latent (as shown in (a)) have
any corresponding pores in the mated full fingerprint.
latents, and automatically extracted from full fingerprints in the background database by using VeriFinger
and the proposed level 3 feature extraction methods. As discussed in Section 3.1, level 3 features are
matched only when the rank 1 minutiae match score of a latent is smaller than the given threshold
(in our experiments, it was empirically set to 80 according to the raw match scores of VeriFinger on
ELFT-EFS-PC, which ranged from 0 to 248). For the sake of efficiency, we matched level 3 features
only between the latent and the top 100 candidate exemplars retrieved by VeriFinger. The match scores
of VeriFinger and level 3 features were then combined by using the weighted sum rule [23]. Before the
fusion, the minutiae match scores of each query latent fingerprint were normalized by using the max-min
normalization method [23] based on the maximum and minimum scores between it and the exemplars.
According to the experimental results, among the 6 latents in ELFT-EFS-PC that have patron pores,
three are already correctly identified at rank 1 by VeriFinger, two (L030G with 91 pores and L201U
with 17 pores, see Figs. 12(a) and 12(b)) are correctly matched with their true mates after rank 1 but
before rank 100, and the remaining one (L014G with 83 pores, see Fig. 3(b)) is correctly matched at rank
2874. By applying the proposed level 3 feature based latent matching method, however, the identification
results of the latents L030G and L201U are not improved. Thisis because i) the number of pores is
small or the pores are sparsely distributed in the latent (e.g. L201U), or ii) the corresponding region of
the latent in its mated full fingerprint image is of poor quality and has few pores (e.g. L030G). Fig. 13
MSU TECHNICAL REPORT, MSU-CSE-10-14, AUGUST 2010 (SUBMITTED TO IEEE TIFS) 24
2 4 6 8 10 12 14 16 18 200.26
0.28
0.3
0.32
0.34
0.36
0.38
0.4
0.42
0.44
RankId
entif
icat
ion
Rat
e
VeriFingerVeriFinger + The Proposed DIP Matcher
Fig. 14. Identification accuracies of VeriFinger and combination of VeriFinger and the proposed
DIP matcher on the ELFT-EFS-PC database.
shows the latent L030G and its mated full fingerprint. Although there are nine mated minutiae in the two
prints shown in Fig. 13, no corresponding pores are found in the full fingerprint for any of the markup
pores in the latent. This is due to the poor quality of the corresponding region in the full fingerprint.
Fig. 14 presents the Cumulative Match Curves of VeriFinger and combination of VeriFinger and the
proposed DIP matcher on the ELFT-EFS-PC database. At rank 1,VeriFinger correctly identified 65
latents; this number was improved to 75 after incorporatingthe proposed DIP matcher. In addition, many
of the other latents had the ranks of their true mates improved. For example, for the latent shown in Fig.
10(b), its true mate was ranked at 82 by VeriFinger; after incorporating the proposed DIP matcher, the
rank was improved to 4. From these results, we can see that level 3 features, when reliably present in
both latents and their mated full fingerprints, are indeed useful in improving the latent matching accuracy.
Unfortunately, there are many difficulties in using level 3 features on the available latent databases. For
one thing, there are very few latents in the latent databaseswith a sufficient number of level 3 features.
This is not only because of the generally low quality of latent fingerprints, but also because latent experts
often are not able to mark some of the level 3 features. While we can clearly see a large number of pores
in the latent shown in Fig. 1(c), according to the markup datain ELFT-EFS-PC, three of the four latent
experts did not mark any pores in it. Another concern is that the quality of the exemplar full fingerprints
in these databases is not good enough to automatically extract reliable level 3 features. As a consequence,
it is not possible to utilize the level 3 features even thoughthey are present in latents (as shown in Fig.
13). Such a problem of reproducibility of level 3 features inexemplar full fingerprints has also been
MSU TECHNICAL REPORT, MSU-CSE-10-14, AUGUST 2010 (SUBMITTED TO IEEE TIFS) 25
2 4 6 8 10 12 14 16 18 200
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Rank
Iden
tific
atio
n R
ate
The Proposed Pore MatcherThe ICP based Pore Matcher
(a)
2 4 6 8 10 12 14 16 18 20
0.65
0.7
0.75
0.8
0.85
0.9
Rank
Iden
tific
atio
n R
ate
VeriFingerVeriFinger + The Proposed Pore MatcherVeriFinger + The ICP based Pore Matcher
(b)
Fig. 15. Identification accuracy of the 131 simulated partial fingerprints and 4180 background
exemplars. (a) Performance of the proposed pore matcher and the ICP based pore matcher. (b)
Performance of combing VeriFinger and pore matchers.
acknowledged in a recent survey on level 3 features among latent examiners [25].
4.5 Simulated Partial Fingerprint Recognition
The experimental results presented above not only show the potential of level 3 features in improving
latent matching accuracy, but also demonstrate the difficulty of using level 3 features in existing latent
databases due to the small number of latents having level 3 features and the poor quality of mated
full fingerprints. In order to better understand the utilityof level 3 features, in particular pores, we
constructed an additional set of 131 simulated partial fingerprint images of small area (320×240 pixels).
They were cropped from the 1000ppi “B” session rolled ink fingerprint images (∼1500×1500 pixels) in
NIST SD30. The pores in these 131 partial fingerprints were manually marked, whereas the minutiae and
ridge skeletons were extracted by VeriFinger [19]. The background database in the experiments was the
same as in our experiments with ELFT-EFS-PC, except that theexemplars from the “B” session rolled
ink fingerprint images in NIST SD30 were substituted with thecorresponding “A” session images in
the database (since the simulated partial fingerprint images as the query fingerprints were cropped from
the “B” session images). All features in the exemplars were automatically extracted by VeriFinger and
the proposed level 3 feature extraction methods. Next, we report the performance of the proposed pore
matching method, and we compare it with VeriFinger and the ICP based pore matching method [10] to
show the effectiveness of the proposed method and the utility of pores.
MSU TECHNICAL REPORT, MSU-CSE-10-14, AUGUST 2010 (SUBMITTED TO IEEE TIFS) 26
2 4 6 8 10 12 14 16 180
50
100
150
200
250
300
350
400
450
Number of MinutiaeM
inut
iae
Mat
ch S
core
s
Fingerprints: Rank=1 by both VeriFinger & ProposedFingerprints: Rank>1 by VeriFinger & Rank=1 by Proposed
Fig. 16. Number of minutiae vs. the minutiae match scores in fingerprints with different
identification results.
The identification accuracy on this dataset is presented in Fig. 15. Using VeriFinger, among the 131
query partial fingerprints, 80 fingerprints were correctly identified at rank 1. After incorporating the
proposed pore matcher, 27 additional fingerprints were successfully identified at rank 1. Compared with
the results on ELFT-EFS-PC, these results are much more promising; they show the effectiveness of level
3 features in matching partial fingerprints of small area which will otherwise pose a challenge to minutiae
based AFIS due to the limited number of minutiae they contain. Moreover, the importance of collecting
good quality full fingerprints can also be seen from these results. With the fast development of fingerprint
imaging techniques and the widespread use of high resolution (1000 ppi) live-scan fingerprint scanners,
we believe that it is becoming feasible to collect good quality fingerprints. This will facilitate extraction
of reliable level 3 features and thereby further improve thelatent matching accuracy of existing AFIS by
incorporating level 3 features.
Fig. 16 compares the fingerprints which are correctly identified at rank 1 by both VeriFinger and the
proposed method with those which are correctly identified after rank 1 by VeriFinger, but at rank 1
after incorporating the proposed method. Interestingly, the improvement due to level 3 features is only
in situations where fingerprints have a small number of minutiae or low minutiae match scores. This
indicates that i) minutiae matchers usually work very well when there is a sufficient number of minutiae,
and ii) the contribution of level 3 features is more effective for fingerprints which have few minutiae
or low minutiae match scores. In [25], about one-third of theparticipant latent examiners reported that
they do not consider level 3 features when level 2 features are of sufficient quality and are sufficient in
number, which confirms our findings here. Moreover, these observations also justify the proposed level
MSU TECHNICAL REPORT, MSU-CSE-10-14, AUGUST 2010 (SUBMITTED TO IEEE TIFS) 27
(a) (b)
Fig. 17. Pore matching results on an example partial fingerprint and its mated full fingerprint
(cropped for display purposes) by using (a) the proposed method and (b) the ICP based method.
The polygons highlight the corresponding reference minutiae between the fingerprints. Due to
the large distortion in the fingerprints, the ICP based method mis-matches many pores; on the
contrary, the proposed method is more robust to distortion, and can still correctly match most of
the pores.
3 feature based latent fingerprint matching algorithm sketched in Section 3.1.
For comparison, the ICP based level 3 feature matching method in [10] has also been implemented.
Given a pair of mated minutiae between two fingerprints, the method in [10] first uses the two mated
minutiae to align the level 3 features (here, pores) in the two fingerprints, and then employs the ICP
algorithm to further align and match the level 3 features. The match score for the two fingerprints was
finally computed based on the average distance,d̄, between the obtained mated pores:sPOR = 1−d̄/dMax,
wheredMax is the maximum distance between mated pores (in our experiments, the distance threshold
between two mated pores was set to 10 pixels). The CMC of the ICP based method in Fig. 15(a) shows
that its performance is much worse than the proposed method:Only 10 fingerprints are correctly identified
at rank 1. By fusing its scores with the scores of VeriFinger,we did not get consistent improvement in
the identification accuracy (see Fig. 15(b)).
Fig. 17 shows the pore matching results on an example partialfingerprint and its mated full fingerprint
(cropped for display purposes) by using the proposed methodand the ICP based method, respectively.
Obvious distortion can be observed between the two fingerprints (note the polygons highlight the corre-
sponding minutiae). As a consequence, in the ICP based method, most pores are falsely matched (see
Fig. 17(b)). On the contrary, most pores are correctly matched by the proposed method despite the large
distortion (see Fig. 17(a)). Based on the match scores, the rank of the mated fingerprint in this pair is 1
MSU TECHNICAL REPORT, MSU-CSE-10-14, AUGUST 2010 (SUBMITTED TO IEEE TIFS) 28
(the proposed method) and 16 (the ICP based method), respectively, while VeriFinger ranks it at 5. These
results illustrate the advantage of the proposed method in considering ridge structures and inter-feature
topology when matching level 3 features in fingerprints.
5 CONCLUSIONS
We have studied the utility of level 3 features for latent fingerprint matching. Automatic algorithms have
been proposed for extracting and matching several level 3 features, i.e. pores, dots, incipient ridges,
and ridge edge protrusions. The proposed matching algorithm compares level 3 features along ridges,
and enforces the topological relationship between level 3 features, minutiae, and ridges. It is thus more
effective for latent fingerprints of small area and robust todistortion (on the set of simulated partial
fingerprints, after combination with VeriFinger minutiae matcher, the rank 1 identification rate by the
proposed method is∼ 81.8%, whereas that by the existing ICP based method is∼ 61.1%). Based on the
performance of the proposed algorithms on operational latent databases and simulated partial fingerprints,
we have shown that
• Published level 3 feature matching algorithms provide poorperformance on operational latent fin-
gerprint databases. Further, conclusions made by previouslevel 3 studies do not hold for challenging
latent fingerprint matching problems;
• The limited number of latents that have level 3 features and the low reproducibility of level 3 features
in poor quality exemplar full fingerprints make the level 3 features of limited use in improving the
latent matching accuracy on the available latent databases;
• The latent matching accuracy can be improved by level 3 features if the features can be reliably
extracted in both latents and mated full fingerprints (the rank 1 accuracy on the ELFT-EFS-PC
database is improved from∼ 26.9% to ∼ 31% after incorporating the proposed DIP matcher);
• The contribution of level 3 features is more effective when the number of minutiae in latents is
small or the minutiae-based match score is low.
While the results of our study demonstrate the potential of level 3 features in improving latent matching
accuracy, they also show the difficulties in using level 3 features. In order to better explore level 3 features
in latent matching and to improve the latent matching accuracy of contemporary AFIS, we would like
to make the following recommendations:
• Additional attention should be paid during enrollment to ensure that a sufficient number of level 3
features in fingerprints be captured. As observed in this study, poor quality of full fingerprints is
one of the major reasons for limited contribution of level 3 features. Using high resolution (≥1000
MSU TECHNICAL REPORT, MSU-CSE-10-14, AUGUST 2010 (SUBMITTED TO IEEE TIFS) 29
ppi) scanners is a necessary but not a sufficient condition for extracting level 3 features. Fingerprint
image quality assessment software that can provide real-time quality level estimates will be very
useful. Since existing fingerprint image quality assessment algorithms (e.g. [22][26]) do not consider
level 3 features, new fingerprint quality assessment algorithms need to be developed;
• Tiny latent fingerprints that have few minutiae but very clear level 3 features should not be simply
discarded. As current AFIS generally cannot correctly identify latents with very few minutiae (say
fewer than 6), more attention should be paid to them in terms of level 3 features, provided that a
large percentage of level 3 features are available in them;
• Latent experts and AFIS developers should cooperate to forma consensus on how to mark level 3
features and ridge quality map in order to optimize the utility of level 3 features. As we observed
in the ELFT-EFS-PC database, level 3 feature marking in manylatents is not satisfactory from the
viewpoint of AFIS developers. In addition, large inconsistency exists among different latent experts.
The close cooperation between latent experts and AFIS developers will in the end help improve the
fingerprint standard, such as CDEFFS [5];
• Use level 3 features only when it is necessary. Marking level3 features in latents is time-consuming
and tedious. Furthermore, when a large number of minutiae are available, the additive value of level
3 features is limited. A simple indicator for using level 3 features is when the minutiae match score
at rank 1 is lower than a predefined threshold;
• ELFT-EFS evaluation would be more useful if the fingerprint image database was made open, and
the utility of level 3 features was separately tested instead of being mixed with other extended
features.
ACKNOWLEDGMENT
This work was supported by NIJ grant 2007-RG-CX-K183. The authors would also like to acknowledge
the support provided by FBI Biometric Center of Excellence during this work.
REFERENCES
[1] H. C. Lee and R. E. Gaensslen (eds.),Advances in Fingerprint Technology, 2nd ed. Elsevier, 2001.
[2] D. Maltoni, D. Maio, A. K. Jain, and S. Prabhakar,Handbook of Fingerprint Recognition, 2nd ed. Springer-Verlag, 2009.
[3] D. R. Ashbaugh,Quantitative-Qualitative Friction Ridge Analysis: An Introduction to Basic and Advanced Ridgeology. CRC
Press, 1999.
[4] A. Roddy and J. Stosz, “Fingerprint features - Statistical analysis and system performance estimates,”Proceedings of the
IEEE, vol. 85(9), pp. 1390-1421, 1997.
MSU TECHNICAL REPORT, MSU-CSE-10-14, AUGUST 2010 (SUBMITTED TO IEEE TIFS) 30
[5] CDEFFS,Data Format for the Interchange of Extended Fingerprint andPalmprint Features, Draft Version 0.5, April 2010,